Codebase for Invariance Through Latent Alignment

Invariance Through Latent Alignment

Takuma Yoneda*, Ge Yang*, Matthew Walter, Bradly Stadie

[Paper] [Website]

ILA-method.webm

Dependency

This codebase requires ml-logger and params_proto. Please look at ila/docker/Dockerfile or Pipfile for more dependencies.

How to use

1. Set Args.checkpoint_root to your local path. This is done by setting $SNAPSHOT_ROOT environment variable
   1. Make sure the path looks like file:///root/subdirectory/subsubdirectory
2. Download pretrained agents
   1. Coming soon
   2. You can also train agents by yourself and store their weights
3. Generate and save trajectories on source (i.e., non-distracted) and target (i.e., distracted) environments
4. Run adapt.py to perform adaptation

If you find our work useful in your research, please consider citing the paper as follows:

@misc{yoneda2021invariance,
      title={Invariance Through Latent Alignment}, 
      author={Takuma Yoneda and Ge Yang and Matthew R. Walter and Bradly Stadie},
      year={2021},
      eprint={2112.08526},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}